TokenMark own benchmark run: NVIDIA H100 PCIe, 2026-10-14

Rented from Lambda Cloud, 1 GPU. Measured by TokenMark, not a vendor or community figure.
nvidia-smi (name, driver, power limit, max SM clock, PCIe gen, PCIe width, memory): NVIDIA H100 PCIe, 570.148.08, 350.00 W, 1755 MHz, 5, 16, 81559 MiB
llama.cpp b11609, built from source with CUDA 12.8 for sm_90 (the rows' "CUDA" setting is the driver's CUDA version).
Benchmark: poor-pauls-benchmark (ppb) 2668a04ce007d8b7aecb2af2e896435cc2506cef, llama-server runner, no extra flags.
Settings: ctx 8192, batch 512 and 1024, 1 concurrent user, 10 ShareGPT prompts x 256 generated tokens.
tok/s = generated tokens / wall time per request, includes time to first token. Each row is the median
of its baseline runs (rows whose runs disagree by more than 25% are dropped), same rule as the ppb rows
from other users on tokenmark.app.

Rows:
  gemma-4-26B-A4B-it (unsloth/gemma-4-26B-A4B-it-GGUF) UD-Q8_K_XL: 138.01 tok/s, TTFT 99.2 ms, 202.1 W, 2 run(s)
  Qwen3.5-27B (unsloth/Qwen3.5-27B-GGUF) Q4_K_M: 52.32 tok/s, TTFT 284.6 ms, 321.1 W, 2 run(s)
  Qwen3.5-35B-A3B (unsloth/Qwen3.5-35B-A3B-GGUF) Q4_K_M: 151.3 tok/s, TTFT 133.5 ms, 184.3 W, 2 run(s)
  Qwen3.6-27B (Jackrong/Qwen3.6-27B-GGUF) Q4_K_M: 52.7 tok/s, TTFT 284.4 ms, 318.2 W, 2 run(s)
  Qwen3.6-27B (Jackrong/Qwen3.6-27B-GGUF) Q8_0: 44.76 tok/s, TTFT 276.5 ms, 322.4 W, 2 run(s)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q4_K_M: 152.64 tok/s, TTFT 140.7 ms, 179.5 W, 2 run(s)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q4_K_XL: 152.1 tok/s, TTFT 128.5 ms, 187.7 W, 2 run(s)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q5_K_M: 151.33 tok/s, TTFT 130.7 ms, 196.5 W, 2 run(s)
  Qwen3.6-35B-A3B (unsloth/Qwen3.6-35B-A3B-GGUF) UD-Q6_K: 146.12 tok/s, TTFT 130.8 ms, 198.6 W, 2 run(s)

results.jsonl: every ppb result row from this run, one per line, in the flat column layout of the
public ppb dataset (made with ppb's own flattener at the SHA above).